|
| 1 | +# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +import argparse |
| 16 | +import time |
| 17 | +import numpy as np |
| 18 | +import os |
| 19 | + |
| 20 | +import paddle |
| 21 | +from paddlenlp.data import Stack, Tuple, Pad |
| 22 | +from paddlenlp.transformers import ErnieTinyTokenizer |
| 23 | +from paddle_serving_client import Client |
| 24 | +from scipy.special import softmax |
| 25 | + |
| 26 | +parser = argparse.ArgumentParser() |
| 27 | +parser.add_argument( |
| 28 | + "--client_config_file", |
| 29 | + type=str, |
| 30 | + default="./serving_client/serving_client_conf.prototxt", |
| 31 | + help="Client prototxt config file.") |
| 32 | +parser.add_argument( |
| 33 | + "--server_ip_port", |
| 34 | + type=str, |
| 35 | + default="127.0.0.1:8090", |
| 36 | + help="The ip address and port of the server.") |
| 37 | +parser.add_argument( |
| 38 | + "--batch_size", |
| 39 | + type=int, |
| 40 | + default=1, |
| 41 | + help="Batch size per GPU/CPU for training.") |
| 42 | +parser.add_argument( |
| 43 | + "--max_seq_length", |
| 44 | + type=int, |
| 45 | + default=128, |
| 46 | + help="The maximum total input sequence length after tokenization. Sequences longer than this will be truncated, sequences shorter will be padded." |
| 47 | +) |
| 48 | +args = parser.parse_args() |
| 49 | + |
| 50 | + |
| 51 | +def convert_example(example, |
| 52 | + tokenizer, |
| 53 | + label_list, |
| 54 | + max_seq_length=512, |
| 55 | + is_test=False): |
| 56 | + """ |
| 57 | + Builds model inputs from a sequence or a pair of sequence for sequence classification tasks |
| 58 | + by concatenating and adding special tokens. And creates a mask from the two sequences passed |
| 59 | + to be used in a sequence-pair classification task. |
| 60 | + |
| 61 | + A BERT sequence has the following format: |
| 62 | +
|
| 63 | + - single sequence: ``[CLS] X [SEP]`` |
| 64 | + - pair of sequences: ``[CLS] A [SEP] B [SEP]`` |
| 65 | +
|
| 66 | + A BERT sequence pair mask has the following format: |
| 67 | + :: |
| 68 | + 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 |
| 69 | + | first sequence | second sequence | |
| 70 | +
|
| 71 | + If only one sequence, only returns the first portion of the mask (0's). |
| 72 | +
|
| 73 | +
|
| 74 | + Args: |
| 75 | + example(obj:`list[str]`): List of input data, containing text and label if it have label. |
| 76 | + tokenizer(obj:`PretrainedTokenizer`): This tokenizer inherits from :class:`~paddlenlp.transformers.PretrainedTokenizer` |
| 77 | + which contains most of the methods. Users should refer to the superclass for more information regarding methods. |
| 78 | + label_list(obj:`list[str]`): All the labels that the data has. |
| 79 | + max_seq_len(obj:`int`): The maximum total input sequence length after tokenization. |
| 80 | + Sequences longer than this will be truncated, sequences shorter will be padded. |
| 81 | + is_test(obj:`False`, defaults to `False`): Whether the example contains label or not. |
| 82 | +
|
| 83 | + Returns: |
| 84 | + input_ids(obj:`list[int]`): The list of token ids. |
| 85 | + token_type_ids(obj: `list[int]`): List of sequence pair mask. |
| 86 | + label(obj:`numpy.array`, data type of int64, optional): The input label if not is_test. |
| 87 | + """ |
| 88 | + text = example |
| 89 | + encoded_inputs = tokenizer(text=text, max_seq_len=max_seq_length) |
| 90 | + input_ids = encoded_inputs["input_ids"] |
| 91 | + token_type_ids = encoded_inputs["token_type_ids"] |
| 92 | + |
| 93 | + if not is_test: |
| 94 | + # create label maps |
| 95 | + label_map = {} |
| 96 | + for (i, l) in enumerate(label_list): |
| 97 | + label_map[l] = i |
| 98 | + |
| 99 | + label = label_map[label] |
| 100 | + label = np.array([label], dtype="int64") |
| 101 | + return input_ids, token_type_ids, label |
| 102 | + else: |
| 103 | + return input_ids, token_type_ids |
| 104 | + |
| 105 | + |
| 106 | +def predict(data, label_map, batch_size): |
| 107 | + """ |
| 108 | + Args: |
| 109 | + sentences (list[str]): each string is a sentence. If have sentences then no need paths |
| 110 | + paths (list[str]): The paths of file which contain sentences. If have paths then no need sentences |
| 111 | + Returns: |
| 112 | + res (list(numpy.ndarray)): The result of sentence, indicate whether each word is replaced, same shape with sentences. |
| 113 | + """ |
| 114 | + |
| 115 | + # initialize client |
| 116 | + client = Client() |
| 117 | + client.load_client_config(args.client_config_file) |
| 118 | + client.connect([args.server_ip_port]) |
| 119 | + |
| 120 | + # TODO: Text tokenization which is done in the serving end not the client end may be better. |
| 121 | + tokenizer = ErnieTinyTokenizer.from_pretrained("ernie-tiny") |
| 122 | + examples = [] |
| 123 | + for text in data: |
| 124 | + input_ids, token_type_ids = convert_example( |
| 125 | + text, |
| 126 | + tokenizer, |
| 127 | + label_list=label_map.values(), |
| 128 | + max_seq_length=args.max_seq_length, |
| 129 | + is_test=True) |
| 130 | + examples.append((input_ids, token_type_ids)) |
| 131 | + |
| 132 | + batchify_fn = lambda samples, fn=Tuple( |
| 133 | + Pad(axis=0, pad_val=tokenizer.pad_token_id, dtype='int64'), # input ids |
| 134 | + Pad(axis=0, pad_val=tokenizer.pad_token_id, dtype='int64'), # token type ids |
| 135 | + ): fn(samples) |
| 136 | + |
| 137 | + # Seperates data into some batches. |
| 138 | + batches = [ |
| 139 | + examples[idx:idx + batch_size] |
| 140 | + for idx in range(0, len(examples), batch_size) |
| 141 | + ] |
| 142 | + |
| 143 | + results = [] |
| 144 | + for batch in batches: |
| 145 | + input_ids, token_type_ids = batchify_fn(batch) |
| 146 | + fetch_map = client.predict( |
| 147 | + feed={"input_ids": input_ids, |
| 148 | + "token_type_ids": token_type_ids}, |
| 149 | + fetch=["save_infer_model/scale_0.tmp_1"], |
| 150 | + batch=True) |
| 151 | + output_data = np.array(fetch_map["save_infer_model/scale_0.tmp_1"]) |
| 152 | + probs = softmax(output_data, axis=1) |
| 153 | + idx = np.argmax(probs, axis=1) |
| 154 | + idx = idx.tolist() |
| 155 | + labels = [label_map[i] for i in idx] |
| 156 | + results.extend(labels) |
| 157 | + |
| 158 | + return results |
| 159 | + |
| 160 | + |
| 161 | +if __name__ == '__main__': |
| 162 | + paddle.enable_static() |
| 163 | + data = [ |
| 164 | + '这个宾馆比较陈旧了，特价的房间也很一般。总体来说一般', |
| 165 | + '怀着十分激动的心情放映，可是看着看着发现，在放映完毕后，出现一集米老鼠的动画片', |
| 166 | + '作为老的四星酒店，房间依然很整洁，相当不错。机场接机服务很好，可以在车上办理入住手续，节省时间。', |
| 167 | + ] |
| 168 | + label_map = {0: 'negative', 1: 'positive'} |
| 169 | + results = predict(data, label_map, args.batch_size) |
| 170 | + for idx, text in enumerate(data): |
| 171 | + print('Data: {} \t Label: {}'.format(text, results[idx])) |
0 commit comments